Abstract Monitoring phytoplankton abundance is essential for understanding ecosystem dynamics and detecting harmful algal blooms (HABs). Traditional microscopy‐based single‐cell counting, while accurate, is time‐consuming and poorly suited to high temporal and spatial resolution monitoring. Automated imaging approaches have therefore been developed to assist cell detection and counting, but they face challenges when dealing with colonial forms. In this study, we investigate transfer learning for automated cell counting in digital images of phytoplankton colonies acquired with a FlowCam system. Several Convolutional Neural Network (CNN) architectures pre‐trained on the ImageNet database were fine‐tuned using annotated datasets of two ecologically relevant colonial taxa commonly observed in the English Channel and the North Sea: Pseudo‐nitzschia and Phaeocystis globosa . Model performance and robustness were evaluated using mean absolute error (MAE) and species‐specific accuracy metrics. Across both taxa, DenseNet121 architecture achieved the best performance, with a Top 2 accuracy up to 99% for P.‐nitzschia and a Top 20% accuracy exceeding 87% for P. globosa . Data augmentation improved model robustness and generalization, particularly for colonies with higher cell numbers. These results demonstrate the ability of deep learning to capture complex spatial patterns and improve counting accuracy across taxa with different morphologies. Beyond methodological performance, this study raises questions about the adequacy of current HAB alert thresholds based on microscopic counts when transitioning to automated imaging systems. Transfer learning provides a robust, fast, and scalable approach that complements traditional monitoring and supports the development of improved environmental assessment strategies.
Wacquet et al. (Sun,) studied this question.